一种快速的频域盲语音分离系统

对真实环境中的混迭语音信号进行盲分离是一个非常困难的任务。许多在仿真信号环境下工作很有效的算法常常不能成功地分离真实环境中录取的混迭语音信号。本文通过对一系列预处理(如解相关),独立分量分析和后处理(解排列和尺度不定性)算法进行优化和有机结合,提出了一种快速的频域盲语音分离系统,有效地实现了对卷积混迭的语音信号的盲分离。特别,文中引入一种改进的数据白化算法,它可以节省运算量并提高独立分量分析算法的收敛速度。基于真实录音信号的仿真实验显示该系统能够快速和稳健地分离卷积混迭的语音信号。

A Fast Frequency Domain Blind Speech Separation System

It is difficult to separate the mixed speech signals recorded in real environment blindly. Many methods, which are effective in simulate environment, cannot separate the mixed speech signals recorded in real environment. This paper describes a fast frequency blind source separation system for separating convolutive mixed audio signals. A set of preprocessing and post-processing methods are combined and optimized to achieve the robust of the entire system. Especially, A symmetrical data whitening approach is proposed, compared with the regular orthogonal whitening method, it can greatly improve the convergence rate of the information maximum independent component analysis algorithm. Experimental results using speech signals recorded in a real room show that the proposed method realizes fast and robust separation for convolutive mixture.